A Novel Technique for the Selection of Speech Segments for Speaker Verification

This paper presents a novel selection technique to discard portions of speech that result in poor discrimination ability in speaker verification tasks. Theory supporting the significance of a sub-segmental selection procedure for test segments, prior to making decisions, is also developed. This approach has the ability to reduce the effect of the regions of the feature space that are not fully modeled by the speaker adaptation algorithm. The proposed technique utilises the frame-based score of the claimed and impostor speakers to select the most discriminative parts of the test segment. The frame selection technique, together with score normalisation, is evaluated on male and female speaker populations separately. Compared with a baseline system using both CMS and variance normalization, the proposed segment selection technique brings 6 % , 18 % relative reductions in error rate for female speakers, while for male speakers a more significant relative error rate reduction of 10%, 20 % is achieved, in terms of EER and minimum DCF respectively. 1